CNTK 序列模型错误:检测到不同的小批量布局

Question

我正在尝试使用 CNTK 训练一个模型，该模型接受两个输入序列并输出二维标量标签。我像这样定义模型：

def create_seq_model(num_tokens):
    with C.default_options(init=C.glorot_uniform()):
        i1 = sequence.input(shape=num_tokens, is_sparse=True, name='i1')
        i2 = sequence.input(shape=num_tokens, is_sparse=True, name='i2')
        s1 = Sequential([Embedding(300), Fold(GRU(64))])(i1)
        s2 = Sequential([Embedding(300), Fold(GRU(64))])(i2)
        combined = splice(s1, s2)
        model = Sequential([Dense(64, activation=sigmoid),
                        Dropout(0.1, seed=42),
                        Dense(2, activation=softmax)])
        return model(combined)

我已将数据转换为 CTF 格式。当我尝试使用以下代码片段(从此处的示例进行非常轻微的修改(来训练它时，我收到一个错误：

def train(reader, model, max_epochs=16):
    criterion = create_criterion_function(model)
    criterion.replace_placeholders({criterion.placeholders[0]: C.input(2, name='labels')})
    epoch_size = 500000
    minibatch_size=128
    lr_per_sample = [0.003]*4+[0.0015]*24+[0.0003]
    lr_per_minibatch= [x*minibatch_size for x in lr_per_sample]
    lr_schedule = learning_rate_schedule(lr_per_minibatch, UnitType.minibatch, epoch_size)
    momentum_as_time_constant = momentum_as_time_constant_schedule(700)
    learner = fsadagrad(criterion.parameters,
                   lr=lr_schedule, momentum=momentum_as_time_constant,
                   gradient_clipping_threshold_per_sample=15,
                   gradient_clipping_with_truncation=True)
    progress_printer = ProgressPrinter(freq=1000, first=10, tag='Training', num_epochs=max_epochs)
    trainer = Trainer(model, criterion, learner, progress_printer)
    log_number_of_parameters(model)
    t = 0
    for epoch in range(max_epochs):
        epoch_end = (epoch+1) * epoch_size
        while(t < epoch_end):
            data = reader.next_minibatch(minibatch_size, input_map={
                criterion.arguments[0]: reader.streams.i1,
                criterion.arguments[1]: reader.streams.i2,
                criterion.arguments[2]: reader.streams.labels
            })
            trainer.train_minibatch(data)
            t += data[criterion.arguments[1]].num_samples 
        trainer.summarize_training_progress()

错误是这样的：

Different minibatch layouts detected (difference in sequence lengths or count or start flags) in data specified for the Function's arguments 'Input('i2', [#, *], [132033])' vs. 'Input('i1', [#, *], [132033])', though these arguments have the same dynamic axes '[*, #]'

我注意到，如果我选择两个输入序列长度相同的示例，那么训练函数就可以工作了。不幸的是，这代表的数据量非常小。处理具有不同数据长度的序列的正确机制是什么？我需要填充输入(类似于 Keras 的 pad_sequence((((吗？

Answer 1

两个序列，i1和i2，被意外地视为具有相同的长度。这是因为 sequence.input(...) 的 sequence_axis 参数的默认值为 default_dynamic_axis() 。解决此问题的一种方法是告诉CNTK，这两个序列的长度不同，方法是为每个序列提供唯一的序列轴，如下所示：

i1_axis = C.Axis.new_unique_dynamic_axis('1')
i2_axis = C.Axis.new_unique_dynamic_axis('2')
i1 = sequence.input(shape=num_tokens, is_sparse=True, sequence_axis=i1_axis, name='i1')
i2 = sequence.input(shape=num_tokens, is_sparse=True, sequence_axis=i2_axis, name='i2')